EP1001372B1

EP1001372B1 - Data processing apparatus and data processing methods

Info

Publication number: EP1001372B1
Application number: EP99308836A
Authority: EP
Inventors: Tetsujiro Kondo; Naoki Kobayashi; Koji Ohta
Original assignee: Sony Corp
Current assignee: Sony Corp
Priority date: 1998-11-09
Filing date: 1999-11-05
Publication date: 2007-10-17
Anticipated expiration: 2019-11-05
Also published as: US20040174930A1; US7088768B2; KR20000035347A; US6728310B1; JP4147647B2; DE69937333T2; KR100712376B1; EP1001372A1; DE69937333D1; JP2000151434A

Description

The present invention relates to data processing apparatus and data processing methods, and more particularly, to a data processing apparatus and a data processing method which may allow noise included in data to be removed more effectively.
In general, data such as transmitted or reproduced image data and sound data includes noise which changes as time elapses. To remove the noise included in the data, there have been known methods in which the average, namely, the whole average, of the whole input data is obtained and in which the average of a part of the input data, which is called a moving average, is obtained.
The method in which the whole average is calculated is effective when the degree of noise included in data, namely, the signal-to-noise ratio (S/N ratio) of the data, is uniform. When the S/N ratio of data varies, however, a portion of the data having a low S/N ratio affects a portion of the data having a high S/N ratio to make it difficult to remove the noise effectively in some cases.
In the method in which the moving average is calculated, since the average of data positioned close to the current input data in the time domain is obtained, the processing result is affected by a change in the S/N ratio of the data. In other words, the processing result has a high S/N ratio for a portion of the data having a high S/N ratio, but the processing result has a low S/N ratio for a portion of the data having a low S/N ratio.
European Patent Application Publication No EP-A-0 821 322 (which corresponds to United States Patent No US-B-6 204 891 ) discloses a data processing apparatus according to the pre-characterising part of claim 1 hereof. Specifically, the input data (samples) are intensity values of pixels of a (temporally spaced) sequence of images. To reduce noise in the images, the apparatus evaluates, for a given (spatial) pixel spatial position, a temporally filtered sample (datum) of the present image by combining a temporally filtered sample of the previous image, multiplied by a first weight, with a noisy sample of the present image multiplied by a second weight which is a function of a probability value concerning absence of a noise peak in the noisy sample. The derivation of the probability value involves taking the difference between the noisy sample of the present image and a spatially smoothed version of that sample obtained from a spatial filter that takes the median intensity of, for example, three successive pixels of the image.
European Patent Application Publication No EP-A-0 280 412 discloses another technique for noise reduction in image signals, in which actual image pixel values are replaced by values determined by a weighting function based on an image data noise variance and a pixel value variance, the mean of neighbouring pixels and mean image noise.
According to a first aspect thereof the present invention provides a data processing apparatus according to claim 1 hereof.
According to a second aspect thereof the present invention provides a data processing method according to claim 9 hereof.
Further aspects of the invention provide a program for carrying out the data processing method and a storage medium for storing the program.
The invention will now be described by way of example with reference to the accompanying drawings, throughout which like parts are referred to by like references, and in which:

Fig. 1 is a block diagram of a self-tuning processing circuit according to an embodiment of the present invention.
Fig. 2A is a flowchart of a processing method used before an algorithm is changed.
Fig. 2B is a flowchart of a processing method used after the algorithm is changed.
Fig. 3A is a flowchart of a processing method used before processing elements are changed.
Fig. 3B is a flowchart of a processing method used after the processing elements are changed.
Fig. 4 is a view showing input learning data used for learning.
Fig. 5 is a block diagram showing an example structure of a noise reduction (NR) processing circuit according to the present invention.
Fig. 6A is a view showing the input data to be processed by the NR processing circuit shown in Fig. 5.
Fig. 6B is a view showing the reliability of the input data to be processed by the NR processing circuit shown in Fig. 5.
Fig. 7 is a block diagram showing an example structure of the NR processing circuit shown in Fig. 5.
Fig. 8 is a block diagram showing an example structure of an input reliability calculation section 12 shown in Fig. 7.
Fig. 9 is a view showing the processing of the input reliability calculation section 12 shown in Fig. 8.
Fig. 10 is a block diagram showing an example structure of an output reliability calculation section 13 shown in Fig. 7.
Fig. 11A is a view showing the processing of input data in the NR processing circuit shown in Fig. 7.
Fig. 11B is a view showing the processing of output data in the NR processing circuit shown in Fig. 7.
Fig. 12 is a flowchart showing the processing of the NR processing circuit shown in Fig. 7.
Fig. 13 is a view showing an input-reliability calculation method.
Fig. 14 is a view showing a simulation result of the NR processing circuit shown in Fig. 7.
Fig. 15A is a view showing a computer into which a program for executing processing related to the present invention is installed.
Fig. 15B is a view showing example storage media for storing the program which executes the processing related to the present invention.
Fig. 15C is a view showing cases in which the program for executing the processing related to the present invention is distributed to the computer through a satellite and a network.
Fig. 16 is a block diagram of a computer into which the program for executing the processing related to the present invention is installed.

An embodiment of the present invention will be described below by referring to the drawings.
Fig. 1 shows an example structure of a self-tuning processing circuit according to an embodiment of the present invention.
In this self-tuning processing circuit, an input signal (for example, digital data) to be processed is processed by the most appropriate processing method, and the result of the processing is output as an output signal.
The self-tuning processing circuit is formed of a real-time learning section 1 and a signal processing section 2.
The real-time learning section 1 receives a learning input signal used for learning the processing method by which the input signal to be processed is processed. The real-time learning section 1 controls the signal processing section 2 such that the processing method is learned in real time according to the learning input signal and the processing method obtained as the result of the learning is used for processing, in order to adaptively process the input signal to be processed so that the output signal is improved as time elapses.
The signal processing section 2 adaptively processes the input signal to be processed by the processing method specified by the real-time learning section 1, and outputs the result of the processing as an output signal.
Therefore, since the signal processing section 2 processes the input signal to be processed by the use of the processing method learned in real time as described above, the input signal to be processed is adaptively processed and thereby the output signal, which is obtained by adaptively processing the input signal to be processed, is improved as time elapses.
More specifically, the processing method by which the signal processing section 2 processes the input signal to be processed is improved by real-time learning in the real-time learning section 1. The signal processing section 2 adaptively processes the input signal to be processed by the use of the processing method specified by the real-time learning section 1. As time elapses or as the amount of input signal processed increases, a so-called learning effect gradually appears, and the output signal output from the signal processing section 2 is improved as time elapses. The improvement of the output signal means that the processing method of the signal processing section 2 is improved as time elapses.
As described above, in the self-tuning processing circuit, the processing method for the input signal to be processed is learned in real time and the processing is executed by the processing method obtained as the result of the learning such that the output signal is improved as time elapses. In other words, as the learning proceeds, the processing appropriate for the input signal to be processed is applied to itself. Therefore, the processing which the self-tuning processing circuit shown in Fig. 1 performs can be called self-tuning processing.
An improvement of the output signal means, for example, an image-quality improvement when the input signal to be processed is an image. More specifically, it means an improvement of the S/N ratio or resolution. When the purpose of the processing in the signal processing section 2 is applying an effect, such as smoothing or edge emphasis, the improvement of the output signal means that the effect greatly appears. Therefore, the improvement of the output signal means that the output signal is changed to that having a state which the user desires.
The real-time learning section 1 learns in real time the processing method used for processing the input signal to be processed, according to the input learning signal input to the section. As shown by dotted lines in Fig. 1, the input learning signal can be the input signal to be processed, input to the signal processing section 2, the output signal output from the signal processing section 2, or further, an auxiliary input signal sent to the self-tuning processing circuit to perform learning of the processing method applied to the input signal to be processed. Two or more of the above signals may be used as input learning signals.
In the self-tuning processing circuit, the processing method of the signal processing section 2 is improved every moment when the real-time learning section 1 learns in real time. Therefore, the self-tuning processing circuit is a so-called change-with-time system. The improvement of the processing method is achieved in two cases: a case in which the algorithm of the processing is changed and a case in which the contents of processing steps constituting the series of processing or those of elements constituting the processing are changed.
When "x" indicates the input signal to be processed and "y" indicates an output signal, for example, the processing in the signal processing section 2 is expressed by an expression of y = f(x) with the use of a function "f( )".
When the expression is set to "f(x) = a₀x² + b₀x + c₀, " for example, the processing of the signal processing section 2 used before the algorithm is changed can be expressed by a flowchart shown in Fig. 2A.
In a step S1, a₀x² is calculated. In a step S2, b₀x is calculated. Then, in a step S3, a₀x²+ b₀x + c₀ is calculated to obtain the output signal "y."
In this case, changing the form of the function "f(x)" corresponds to changing the algorithm of the processing. More specifically, changing the function from "f(x) = a₀x² + b₀x + c₀" to "f(x) = (a₀x² + b₀x + c₀)²," for example, corresponds to changing the algorithm of the processing.
In this case, the processing of the signal processing section 2 used after the algorithm is changed can be expressed, for example, by a flowchart shown in Fig. 2B.
Specifically, in steps S1 to S3, the same processing as that shown in Fig. 2A is executed, and (a₀x² + b₀x + c₀)² is calculated to obtain the output signal "y."
As clearly understood from the comparison between the flowchart shown in Fig. 2A and that shown in Fig. 2B, when the algorithm of the processing is changed, the flowchart which expresses the processing is also changed accordingly.
Changing a coefficient of the function "f(x)" corresponds to changing the contents of an element constituting the processing. Specifically, for example, changing the function from "f(x) = a₀x² + b₀x + c₀" to "f(x) = a₁x² + b₁x + c₁" corresponds to changing the contents of the elements constituting the processing.
In this case, the processing of the signal processing section 2 is changed from the processing shown by a flowchart in Fig. 3A, which is the same as that shown in Fig. 2A, to, for example, the processing shown by a flowchart in Fig. 3B.
Specifically, in this case, a₁x² is calculated in a step S1, b₁x is calculated in a step S2, and then, a₁x² + b₁x + c₁ is calculated in a step S3 to obtain the output signal "y."
As clearly understood from the comparison between the flowchart shown in Fig. 3A and that shown in Fig. 3B, when the coefficients of the function "f(x)" are changed, the contents of steps constituting the processing are changed accordingly. Specifically, in Figs. 3A and 3B, the contents of the step S1 is changed from the calculation of a₀x² to that of a₁x², the contents of the step S2 is changed from the calculation of b₀x to that of b₁x, the contents of the step S3 is changed from a₀x² + b₀x + c₀ to a₁x² + b₁x + c₁.
In some cases, the processing method is improved both by changing the algorithm of the processing and by changing the contents of an element constituting the processing.
The functions used in Figs. 2A and 2B and Figs. 3A and 3B are just examples used to describe an improvement of the processing method. Whether the output signal is actually improved is irrelevant to the fact that the function expressing the processing of the signal processing section 2 is changed as described above.
The self-tuning processing circuit differs, for example, from conventional digital filters in that the coefficients of the function f(x) are changed by real-time learning, namely, according to newly generated adaptive coefficients. Specifically, in the conventional digital filters, a tap coefficient may be changed according to input data to a coefficient prepared in advance, which is not generated by real-time learning. In other words, in the conventional digital filters, a tap coefficient which the system designer thinks is appropriate for input data is selected among tap coefficients prepared in advance. A tap coefficient adaptive to input data is not obtained. As a result, even if a tap coefficient more appropriate for the input data exists, it cannot be used unless it is prepared in advance. Therefore, the above conventional digital filters are called in some cases systems changing a tap coefficient "adaptively," but accurately speaking, it is not "adaptively" but "selectively." They are systems in which a tap coefficient adaptive to input data is not generated but a tap coefficient selected uniquely against the input data is used.
In contrast, in the self-tuning processing circuit, a tap coefficient is generated by real-time learning such that input data is adaptively processed. In other words, the input data is processed with the most appropriate tap coefficient. More specifically, for example, whereas in the conventional digital filters, for example, when two sets of tap coefficients are prepared in advance, one of the two sets of tap coefficients which the system designer thinks is appropriate for the input data is uniquely selected and used. Tap coefficients used for processing the input data are improved every moment by real-time learning to obtain tap coefficients more appropriate for the input data, and processing is executed with the use of such tap coefficients in the self-tuning processing circuit. Therefore, processing adaptive to the input data is executed.
In addition, the self-tuning processing circuit differs from conventional processing which applies adaptive processing to input data but output data is not improved every moment, such as adaptive dynamic range coding (ADRC) processing, in that output data is improved every moment. Specifically, in the ADRC processing, the maximum value MAX and the minimum value MIN of the pixel values of pixels constituting a block in an image are detected, and the pixel values of the pixels constituting the block are re-quantized in K bits according to a local dynamic range DR obtained by MAX - MIN in the block. More specifically, the minimum value MIN is subtracted from the value of each pixel in the block and the difference is divided by DR/2^K for requantization. Therefore, adaptive processing is applied to input data in that the quantization step DR/2^K of the pixel values is obtained according to the block dynamic range DR and the pixel values are quantized by the quantization step DR/2^K, but output data obtained as the quantization results is not improved every moment.
The real-time learning section 1 uses all the sample values of input learning signals obtained so far in order to improve the output signal as time elapses in real-time learning at the current time "t." The output signal of the signal processing section 2 may be more effectively improved if the real-time learning section 1 gives a weight to sample values of input learning signals obtained at positions close in a time manner to the current time "t" and sample values obtained at positions far from the current time are not used much. Therefore, the real-time learning section 1 can execute real-time learning such that the above-described weight is given to input learning signals, namely, the sample values of input learning signals obtained at positions in a time manner far from the current time "t" are ignored according to the rates corresponding to the degrees of the time differences. Alternatively, the real-time learning section 1 can use only input learning signals obtained within a predetermined time period from the current time "t." Specifically, for example, as shown in Fig. 4, the real-time learning section 1 is configured such that only input learning signals within four samples from the current time "t" is used for real-time learning at the current time "t." This can be implemented by applying a weight of zero to input learning signals away from the current time "t" by five or more sample periods.
In real-time learning at the current time "t," the real-time learning section 1 can use only sample values within a predetermined range in magnitude from the sample value of the input learning signal obtained at the current time "t." Specifically, for example, in Fig. 4, a sample value, such as the sample value of the input learning signal obtained at the time "t-2," having a magnitude largely different from that of the sample value of the input learning signal obtained at the current time "t" may adversely affect learning. Therefore, it is possible not to use, namely, to give a weight of zero to, a sample value having a magnitude largely different from that of the input learning signal obtained at the current time "t" in real-time learning at the current time "t," in the real-time learning section 1.
Fig. 5 shows an example structure of a noise reduction (NR) processing circuit 100 to which the self-tuning processing circuit shown in Fig. 1 is applied.
In this NR processing circuit 100, when data having noise is input to the NR processing circuit 100, self-tuning processing is applied to the input data and data obtained by effectively removing the noise from the input data is output.
Specifically, for example, to simplify a description, a case is examined in which input data having a constant true value and on which noise fluctuating in time is superposed, as shown in Fig. 6A, is averaged to remove the noise fluctuating in time. The noise is effectively removed by setting a weight for input data having a large noise level, namely, having a low S/N ratio, to a small value and by setting a weight for input data having a small noise level, namely, having a high S/N ratio, to a large value.
In the NR processing circuit 100, as the evaluation value of input data, the degree of reliability indicating how the input data is close to the true value, which is the reliability of the input data against the true value, for example, shown in Fig. 6B, is obtained by real-time learning. The NR processing circuit 100 calculates its average while the weight corresponding to the reliability is applied to the input data, to effectively remove the noise.
Therefore, the NR processing circuit 100 obtains a weighted average of the input data by the use of the weight corresponding to its reliability and outputs it as output data. The output data y(t) is obtained from the following expression, where x(t), y(t), and α_x(t) indicate the input data, the output data, and the reliability of the input data at the time "t" in the NR processing circuit 100, respectively. $y (t) = \frac{\sum_{i = 0}^{t} α_{x (i)} x (i)}{\sum_{i = 0}^{t} α_{x (i)}}$
It is assumed here that a larger weight is given to the higher degree of reliability α_x(t) of input data.
The output data "y(t-1)," obtained one sample before the current time "t," is calculated as follows from the expression (1). $y (t - 1) = \frac{\sum_{i = 0}^{t - 1} α_{x (i)} x (i)}{\sum_{i = 0}^{t - 1} α_{x (i)}}$
As for the output data y(t), the degree of reliability α_y(t) indicating how the output data y(t) is close to the true value, which is the reliability of the output data y(t) against the true value, is introduced as an evaluation value of the output data y(t). The reliability α_y(t-1) of the output data y(t-1), obtained one sample before the current time "t," is defined by the following expression. $α_{y (t - 1)} = \sum_{i = 0}^{t - 1} α_{x (i)}$
In this case, the output data "y(t)" and its reliability α_y(t) are expressed as follows with the use of the expressions (1) to (3). $\begin{matrix} y (t) = \frac{\overset{α_{y (t - 1)} y (t - 1)}{\overset{︷}{\sum_{i = 0}^{t} α_{x (i)} x (i)}} α_{x (t)} x (t)}{\underset{α_{y (t - 1)}}{\underset{︸}{\sum_{i = 0}^{t - 1} α_{x (i)}}} + α_{x (t)}} \\ = \frac{α_{y (t - 1)} y (t - 1) + α_{x (t)} x (t)}{α_{y (t - 1)} + α_{x (t)}} \\ α_{y (t)} = α_{y (t - 1)} + α_{x (t)} . \end{matrix}$
The weight used to obtain the output data y(t) at the time "t" is indicated by w(t), and defined by the following expression. $w (t) = α_{y (t - 1)} / (α_{y (t - 1)} + α_{x (t)})$
From the expression (5), the following expression is satisfied. $1 - w (t) = α_{x (z)} / (α_{y (z - 1)} + α_{x (t)})$
With the use of the expressions (5) and (6), the output data y(t) in the expression (4) can be expressed in the following way by a weighted average obtained by multiplications and an addition. $y (t) = w (t) y (t - 1) + (1 - w (t)) x (t)$
The weights w(t) and 1 - w(t) used in the expression (7) can be obtained from the expression (5) with the use of the reliability α_y(t-1) of the output data y(t-1) obtained one sample before and the reliability α_x(t) of the current input data x(t). The reliability α_y(t) of the current output data y(t) in the expression (4) can also be obtained with the use of the reliability α_y(t-1) of the output data y(t-1) obtained one sample before and the reliability α_x(t) of the current input data x(t).
When the reliability α_x(t) of the input data x(t) and the reliability α_y(t) of the output data y(t) are expressed, respectively, by the reciprocals of the corresponding dispersions σ_x(t) ² or σ_y(t) ², in other words, when the reliability α_x(t) and the reliability α_y(t) are set as follows, $\begin{matrix} α_{x (t)} = 1 / {σ_{x (t) ()}}^{2} \\ α_{y (t)} = 1 / {σ_{y (t) ()}}^{2} \end{matrix}$
the weight w(t) in the expression (7) can be obtained by the following expression. $w (t) = {σ_{x (t) ()}}^{2} / ({σ_{y (t - 1)}}^{2} + {σ_{x (t)}}^{2})$
In this case, 1 - w(t) in the expression (7) can be obtained by the following expression. $1 - w (t) = {σ_{y (t - 1)}}^{2} / ({σ_{y (t - 1)}}^{2} + {σ_{x (t)}}^{2})$
A term, σ_y(t) ², can be obtained by the following expression. ${σ_{y (t) ()}}^{2} = w {(t)}^{2} {σ_{y (t - 1)}}^{2} + {(1 - w (t) ())}^{2} {σ_{x (t)}}^{2}$
The NR processing circuit 100 shown in Fig. 5 learns in real time the weight w(t) according to the expression (5) and 1 - w(t) according to the expression (6), with the input data x(t) and the output data y(t) being used as input learning signals. The NR processing circuit calculates the weighted average of the output data y(t-1) obtained one sample before and the current input data x(t) according to the expression (7) with the use of the weight w(t) obtained as a result of the real-time learning to adaptively process the input data x(t) so that the noise included in the input data x(t) is effectively removed. In other words, the S/N ratio of the output data y(t) output from the NR processing circuit 100 is improved as time elapses.
Fig. 7 shows an example structure of such an NR processing circuit 100.
A latch circuit 11₁ receives data. The latch circuit 11₁ latches the input data, for example, in synchronization with the timing when the input data is received, and sends it to a subsequent-stage latch circuit 11₂ and to an input reliability calculation section 12. The latch circuit 11₂ and a latch circuit 11₃ latch the data output from the previous-stage latch circuits 11₁ and 11₂, and send it to the subsequent-stage latch circuits 11₃ and 11₄, respectively, and to the input reliability calculation section 12, in the same way as the latch circuit 11₁. The latch circuit 11₄ latches the data output from the previous-stage latch circuit 11₃ and sends it to the input reliability calculation section 12.
The input reliability calculation section 12 receives the same input data as that sent to the latch circuit 11₁ in addition to the input data latched by the latch circuits 11₁ to 11₄. Therefore, when input data x(t) is sent to the latch circuit 11₁ and to the input reliability calculation section 12, the input reliability calculation section 12 also receives input data x(t-1) to x(t-4) latched by the latch circuits 11₁ to 11₄. The input reliability calculation circuit 12 calculates, for example, the dispersion of input data from the input data x(t) to x(t-4), and sends the reciprocal of the dispersion as the reliability α_x(t), namely, the input reliability, of the input data x(t) to an output reliability calculation section 13 and to a weight calculation section 15.
The output reliability calculation section 13 obtains the reliability α_y(t), namely, the output reliability of the output data y(t) according to the expression (4) with the use of the input reliability α_x(t) sent from the input reliability calculation section 12 and the output of a latch circuit 14, and outputs it to the latch circuit 14.
The latch circuit 14 latches the output reliability α_y(t) sent from the output reliability calculation section 13, for example, in synchronization with latching of the input data x(t), and sends it to the output reliability calculation section 13 and the weight calculation section 15. Therefore, the latch circuit 14 sends the output reliability α_y(t-1) obtained one sample before, to the output reliability calculation section 13 and the weight calculation section 15.
The weight calculation section 15 obtains a weight w(t) according to the expression (5) with the use of the input reliability α_x(t) sent from the input reliability calculation section 12 and the output reliability α_y(t-1) sent from the latch circuit 14, and sends it to a weighting section 21 and an operation section 22.
The latch circuits 11₁ to 11₄, the input reliability calculation section 12, the output reliability calculation section 13, the latch circuit 14, and the weight calculation section 15, described above, correspond to the real-time learning section shown in Fig. 1.
The weighting section 21 multiplies the weight w(t) sent from the weight calculation section 15 by the output of a latch circuit 25, and sends the product to an operation section 24. The operation section 22 subtracts the weight w(t) sent from the weight calculation section 15, from one, and sends the subtraction result, 1 - w(t), to a weighting section 23. The weighting section 23 receives the input data x(t) as well as the output of the operation section 22. The weighting section 23 multiplies the input data x(t) by the output of the operation section 22, and sends the product to the operation section 24. The operation section 24 adds the outputs of the weighting sections 21 and 23, outputs the sum as output data y(t), and sends it to the latch circuit 25. The latch circuit 25 latches the output data of the operation section 24, for example, in synchronization with latching of the input data x(t), and sends it to the weighting section 21.
The weighting section 21, the operation section 22, the weighting section 23, the operation section 24, and the latch circuit 25, described above, correspond to the signal processing section 2 shown in Fig. 1, and calculate the weighted average according to the expression (7) and outputs it as the output data y(t).
Fig. 8 shows an example structure of the input reliability calculation section 12 shown in Fig. 7.
As described above, the input reliability calculation section 12 receives the current input data x(t) and, in addition, the input data x(t-1) to x(t-4), up to four samples before the current one. The input reliability calculation section 12 obtains the dispersion of the five-sample input data x(t) to x(t-4) as shown in Fig. 9, and outputs the reciprocal of the dispersion as the input reliability α_x(t).
Specifically, the input data x(t) to x(t-4) is sent to an average calculation circuit 31 and to a dispersion calculation circuit 32. The average calculation circuit 31 calculates the average m(t) of the five-sample input data x(t) to x(t-4) according to the following expression and sends the result to the dispersion calculation circuit 32. $m (t) = \frac{\sum_{n = 0}^{N - 1} x (t - n)}{N}$
In the embodiment shown in Fig. 8, N equals 5 in the expression (12).
The dispersion calculation circuit 32 uses the input data x(t) to x(t-4) input thereto and the average m(t) sent from the average calculation circuit 31 to calculate the dispersion σ_x(t) ² according to an expression (13) and sends it to a reciprocal calculation circuit 33. $σ_{x (t)}^{2} = \frac{\sum_{i = 0}^{N - 1} {(x (t - n) - m (t))}^{2}}{N}$
In the same way as in the expression (12), N also equals 5 in the expression (13) in the embodiment shown in Fig. 8.
The reciprocal calculation circuit 33 obtains the reciprocal of the dispersion σ_x(t) ² sent from the dispersion calculation circuit 32, and outputs it as the input reliability α_x(t). $α_{x (t)} = \frac{1}{σ_{x (t)}^{2}}$
Fig. 10 shows an example structure of the output reliability calculation section 13 shown in Fig. 7.
As shown in the figure, the output reliability calculation section 13 is formed of an operation section 41. The operation section 41 receives the current input reliability α_x(t) sent from the input reliability calculation section 12 and the output reliability α_y(t-1) obtained one sample before, sent from the latch circuit 14. The operation section 41 adds the input reliability α_x(t) and the output reliability α_y(t-1) according to the expression (4) and outputs the sum as the current output reliability α_y(t).
The operation of the NR processing circuit 100 shown in Fig. 7 will be described below by referring to Figs. 11A and 11B and Fig. 12. The operation of the NR processing circuit 100 and the control of each functional block are implemented by a controller not shown. Alternatively, the whole NR processing algorithm may be implemented by software.
As shown in Fig. 11A, the NR processing circuit 100 uses five samples, the current input data x(t) and the input data x(t-1) to x(t-4) obtained up to four samples before the current one, to obtain the dispersion σ_x(t) ² thereof, and further obtains the reciprocal of the dispersion as input reliability α_x(t).
The weight w(t) as well as the reliability α_y(t) of the current output data is obtained as shown in Fig. 11B with the use of the input reliability α_x(t) and the output reliability α_y(t-1) obtained one sample before, and the weighted average of the input data x(t) and the output data y(t-1) obtained one sample before is calculated according to the weight w(t) to output the weighted average as the output data y(t). With the above processing, α_y(t) and σ_x(t) ² are obtained in this order as shown in Fig. 11B.
More specifically, as shown in a flowchart of Fig. 12, the input data x(t) is input to the latch circuit 11₁, the input reliability calculation section 12, and the weighting section 23 in a step S11.
In the next step S12, the input reliability calculation section 12 obtains the input reliability α_x(t).
More specifically, the latch circuit 111 latches the received input data in synchronization with the timing when the input data is received, and sends it to the subsequent-stage latch circuit 11₂ and to the input reliability calculation section 12. The latch circuit 11₂ and the latch circuit 11₃ latch the input data output from the previous-stage latch circuits 11₁ and 11₂ and send the data to the subsequent-stage latch circuits 11₃ and 11₄ and to the input reliability calculation section 12, respectively, in the same way as the latch circuit 11₁. The latch circuit 11₄ latches the input data output from the previous-stage latch circuit 11₃ and sends it to the input reliability calculation section 12. Therefore, the input reliability calculation section 12 receives the input data x(t-1) to x(t-4) from the latch circuits 11₁ to 11₄, respectively, at the same time when it receives the input data x(t). The input reliability calculation section 12 uses the input data x(t) to x(t-4) to obtain the input reliability α_x(t) and sends it to the output reliability calculation section 13 and to the weight calculation section 15, as described above.
At the timing when the input reliability calculation section 12 sends the input reliability α_x(t) to the weight calculation section 15, the latch circuit 14 latches the output reliability α_y(t-1) output from the output reliability calculation section 13 one sample before. In a step S13, the weight calculation section 15 uses the input reliability α_x(t) sent from the input reliability calculation section 12 and the output reliability α_y(t-1) latched by the latch circuit 14 to obtain the weight w(t) according to the expression (5). This weight w(t) is sent to the weighting section 21 and the operation section 22.
The weighting section 21, the operation section 22, the weighting section 23, the operation section 24, and the latch circuit 25 use the weight w(t) output from the weight calculation section 15 to calculate the weighted average of the input data x(t) and the output data y(t-1) obtained one sample before, according to the expression (7).
Specifically, the weighting section 21 multiplies the output of the latch circuit 25 by the weight w(t) sent from the weight calculation section 15, and sends the product to the operation section 24. The latch circuit 25 latches the output data y(t-1) which the operation section 24 outputs the last time, at the timing when the weight calculation section 15 outputs the weight w(t). Therefore, the weighting section 21 obtains the product w(t)y(t-1) of the output data y(t-1) and the weight w(t) and sends it to the operation section 24.
The operation section 22 subtracts the weight w(t) sent from the weight calculation section 15, from one, and sends the subtraction value, 1 - w(t), to the weighting section 23. The weighting section 23 multiplies the output, 1 - w(t), of the operation section 22 by the input data x(t) and sends the multiplication result, (1 - w(t))x(t), to the operation section 24.
The operation section 24 adds the output, w(t)y(t-1), of the weighting section 21 and the output, (1-w(t))x(t), of the weighting section 23. In other words, with the use of the weight w(t) output from the weight calculation section 15, the weighted average of the input data x(t) and the output data y(t-1) obtained one sample before is calculated according to the expression (7).
This weighted average is output as the output data y(t) in a step S15. The output data y(t) is sent to the latch circuit 25 and latched.
In the next step S16, the controller determines whether input data still exists. When it is determined in the step S16 that input data to be processed still exists, the procedure proceeds to a step S17 and the output reliability calculation section 13 updates the output reliability. Specifically, the output reliability calculation section 13 adds the input reliability α_x(t) calculated in the step 512 by the input reliability calculation section 12 and the output reliability α_y(t-1) obtained one sample before and latched by the latch circuit 14, according to the expression (4) to obtain the current output reliability α_y(t), and sends it to the latch circuit 14. Then, the procedure returns to the step S11, and the same processing is repeated with the next input data.
On the other hand, when the controller determines in the step S16 that input data to be processed does not exist, the processing is terminated.
As described above, the input reliability α_x(t) of the current input data x(t) and the output reliability α_y(t-1) obtained one sample before are added to calculate the weight w(t). Then, according to the weight w(t), the weighted average of the current input data x(t) and the output data y(t-1) obtained one sample before is calculated, and the average is output as the output data y(t) serving as the processing result of the input data x(t). The output reliability α_y(t) of the output data y(t) is then obtained by adding the current input reliability α_x(t) to the output reliability α_y(t-1) obtained one sample before. In the same way, input data x(t+1), x(t+2), ... is sequentially processed. Therefore, the weight w(t) is learned such that portions having high noise levels are not much taken into account and portions having low noise levels are sufficiently taken into account in previous input data. In other words, the NR processing circuit 100 obtains a weight w(t) adaptive to input data, and as a result, improves the output data every moment as learning of the weight w(t) proceeds. The NR processing circuit 100 effectively removes noise from the input data and outputs data.
Since the NR processing circuit 100 uses the five-sample input data x(t) to x(t-4) to obtain the dispersion σ_x(t) ² in the embodiment shown in Fig. 7, the input reliability α_x(t), which is the reciprocal of the dispersion, cannot be obtained until the time required for obtaining five samples elapses from the start of the data input. Until the time required for obtaining five samples elapses, the NR processing circuit 100 may execute the processing, for example, in a way in which neither input reliability nor output reliability is calculated and the simple average of input data which has been input is obtained and output as output data. A processing method used immediately after the data input is started is not limited to this method.
In the above-described case, since the NR processing circuit 100 uses the reciprocal of the dispersion of input data as the input reliability α_x(t), the input reliability α_x(t) indicates the scatter of the input data within a predetermined time range. Therefore, when the degree of noise included in input data, for example, the S/N ratio of the input data, varies, the NR processing circuit 100 very effectively removes the noise.
When the reciprocal of the dispersion of the input data is used as the input reliability α_x(t) in the NR processing circuit 100, however, the effect of noise removal is slightly lowered for a local change of a noise level, namely, a change in a very narrow area, due to the nature of the dispersion.
The local change of a noise level can be effectively handled by using, for example, the reciprocal of the error square of the current input data against the average of input data, as the input reliability α_x(t).
In this case, the input reliability α_x(t) is calculated in the following way.
Specifically, for example, the NR processing circuit 100 calculates the average m(t) of five-sample input data x(t) to x(t-4) according to an expression (15) as shown in Fig. 13. $m (t) = \frac{\sum_{i = 0}^{N - 1} x (t - n)}{N}$
In an embodiment shown in Fig. 13, N equals 5 in the expression (15).
With the use of the input data x(t) and the average m(t), the NR processing circuit 100 calculates an error square d_x(t) ² of the input data x(t) against the average m(t) according to an expression (16). $d_{x (t) ()}^{2} = {(x (t) - m (t) ())}^{2}$
Then, the reciprocal of the error square d_x(t) ² is obtained according to the following expression to obtain the input reliability α_x(t). $α_{x (t)} = \frac{1}{d_{x (t)}^{()}}$
As described above, the method using the reciprocal of the error square as the input reliability α_x(t) is especially effective when the average m(t) is close to the true value.
In the above description, the input reliability α_x(t) is obtained based on either the dispersion σ_x(t) ² or the error square d_x(t) ². It can be also obtained based on both. Specifically, for example, the sum of the reciprocal of the dispersion σ_x(t) ² and the reciprocal of the error square d_x(t) ² can be used as an input reliability α_x(t).
Since the dispersion σ_x(t) ² indicates the degree of the local scatters of input data in a rather wide area and the error square d_x(t) ² indicates the degree of the local scatters of the data in a narrow area, when a combination of these factors is used as the input reliability α_x(t), even if the input data has a changing S/N ratio in a rather wide range and a local level of noise included in the input data also changes, the noise is effectively removed.
Fig. 14 shows a simulation result of self-tuning processing in which image input data is processed with its dispersion σ_x(t) ² being used as an input reliability α_x(t).
In Fig. 14, the horizontal axis and the vertical axis indicate a frame count or a frame number and an S/N ratio, respectively. The S/N ratios of input image data are indicated by 0 marks. The S/N ratios of output data obtained by the moving-average method are indicated by + marks, and the S/N ratios of output data obtained by a method according to the present invention are indicated by x marks. Five samples are used to calculate the moving average and the input reliability α_x(t).
As clearly understood from Fig. 14, the S/N ratio of the output data obtained by the moving-average method tracks the S/N ratio of the input image data. Therefore, when the S/N ratio of the image data increases, that of the output data also increases, and when the S/N ratio of the image data decreases, that of the output data also decreases. In contrast, the S/N ratio of the output data obtained by the method according to the present invention increases every moment due to the effect of learning of the weight w(t), and it is almost not subject to a change in the S/N ratio of the image data.
The present invention can also be applied to a case in which noise is removed, for example, from sound data, in addition to image data.
The above embodiment has been described in the viewpoint of noise reduction. Since the NR processing circuit 100 learns a processing method, the weight coefficient w(t) in this case, in real time and executes the processing by the processing method obtained as a result of the learning, such that input data is adaptively processed to improve output data as time elapses, the NR processing circuit 100 has a nature of self-tuning processing in which the input data itself is adaptively processed. In other words, the NR processing circuit can execute, for example, most-appropriate waveform shaping of input data.
In the present embodiment, the output data y(t-1) is multiplied by the weight w(t) and the input data x(t) is multiplied by the weight, 1 - w(t), in the expression (7) to obtain the output data y(t) in the NR processing circuit 100. It is also possible that either the output data y(t-1) or the input data x(t) is multiplied by a weight to obtain an output data y(t).
The series of processing shown in Fig. 12 can be implemented by hardware or software. When the series of processing is achieved by software, a program constituting the software is installed into a computer which is built in the NR processing circuit 100 serving as special hardware or into a general-purpose computer.
Recording media which store the program for executing the series of processing described above will be described below by referring to Figs. 15A, 15B, and 15C. The program is installed into a computer and is made executable by the computer.
The program is recorded in advance into a hard disk 102 or a semiconductor memory 103 serving as a recording medium which is built in a computer 101, as shown in Fig. 15A.
Alternatively, the program is recorded temporarily or perpetually into recording media, such as a floppy disk 111, a compact disc read-only memory (CD-ROM) 112, a magnetooptical (MO) disk 113, a digital versatile disk (DVD) 114, a magnetic disk 115, and a semiconductor memory 116, as shown in Fig. 15B.
The program is installed in the computer from the above-described recording medium. Alternatively, the program is transferred by radio from a downloading site 121 to the computer 101 through an artificial satellite for digital satellite broadcasting, or to the computer 101 by wire through a network 131 such as a local area network (LAN) or the Internet; and is installed into the hard disk 102, built in the computer 101, as shown in Fig. 15C.
In the present specification, the steps describing the programs for executing various types of processing are not necessarily processed in a time sequential manner in the order in which the steps are described in the flowcharts. Processing to be executed in parallel or individually, such as parallel processing or processing by objects, is also included.
Fig. 16 shows an example structure of the computer 101 shown in Figs. 15A, 15B, and 15C.
The computer 101 includes a central processing unit (CPU) 142, as shown in Fig. 16. The CPU 142 is connected to an input and output interface 145 through a bus 141. When the user operates an input section 147 formed of a keyboard and a mouse to input a command through the input and output interface 145, the CPU 142 executes a program stored in a read-only memory (ROM) 143 corresponding to the semiconductor memory 103 shown in Fig. 15A. Alternatively, the CPU 142 loads into a random access memory (RAM) 144 a program stored in the hard disk 102; a program transferred through the satellite 122 or the network 131, received by a communication section 148, and installed into the hard disk 102; or a program read from the floppy disk 111, the CD-ROM 112, the MO disk 113, the DVD 114, or the magnetic disk 115 which is loaded into a drive 149, and installed into the hard disk 102; and executes it. The CPU 142 outputs the processing result, for example, through the input and output interface 145 to a display section 146 formed of a liquid-crystal display (LCD), as required.

Claims

A data processing apparatus for processing successive input data and outputting the processed data, the apparatus comprising data processing means (2) for processing the input data by a processing method and outputting the processed data, and the apparatus being characterised by:
input-data evaluation means (12) for calculating a degree of reliability (α_x) of the input data, for each successive input datum, from a group of successive input data;

output-data evaluation means (13) for calculating a degree of reliability (α_y) of the output data for each datum; and

real-time learning means (15) for calculating, for each datum, a processing method parameter (w(t)) indicative of a processing method from the reliability calculations performed, for each datum, by said input-data evaluation means (12) and said output-data evaluation means (13), and supplying said processing method parameter to said data processing means (2) so that it processes the input data according to a learned processing method indicated by the processing method parameter.
A data processing apparatus according to claim 1, comprising input-data storage means (11) for storing a predetermined number of successive input data so as to provide said group of successive input data for said input-data evaluation means (12).
A data processing apparatus according to claim 2, wherein said input-data evaluation means (12) comprises means (32) for calculating the dispersion of the predetermined number of input data stored by said input-data storage means (11), and means (33) for calculating the degree of reliability (α_x) of the input data by forming the reciprocal of the dispersion.
A data processing apparatus according to claim 2, wherein said input-data evaluation means (12) is operative to calculate the average of the predetermined number of input data stored by said input-data storage means (11), and to calculate the degree of reliability (α_x) of the input data according to the error of each datum against the average.
A data processing apparatus according to claim 2, wherein said input-data evaluation means (12) is operative to calculate the dispersion and the average of the predetermined number of input data stored by said input-data storage means (11), to obtain the error of each input datum against the average, and to calculate the degree of reliability of the input data according to the dispersion and the error.
A data processing apparatus according to any one of the preceding claims, wherein said output-data evaluation means (13, 41) is operative to calculate the current output data reliability according to the sum of the calculation of the current input data reliability and the calculation of the output data reliability obtained one datum before.
A data processing apparatus according to any one of the preceding claims, comprising output-data storage means (25) for storing the output data, wherein said data processing means (2) is operative to add the previous output datum stored by said output-data storage means and the current input datum to obtain the output datum corresponding to the current input datum.
A data processing apparatus according to claim 7, wherein said real-time learning means (15) is operative to calculate a processing method parameter that is a predetermined weight coefficient (w(t)), and said data processing means (2) is operative such that its output data corresponds to the current input data according to the weight coefficient.
A data processing method for processing successive input data and outputting the processed data, the method comprising a data processing step (2) for processing the input data by a processing method and outputting the processed data, and the method being characterised by;
an input-data evaluation step (12) for calculating a degree of reliability (α_x) of the input data, for each successive input datum, from a group of successive input data;

an output-data evaluation step (13) for calculating a degree of reliability (α_y) of the output data for each datum; and

a real-time learning step (15) of calculating, for each datum, a processing method parameter (w(t)) indicative of a processing method from the reliability calculations performed, for each datum, in said input-data evaluation step (12) and said output-data evaluation step (13), and using said processing method parameter to process the input data in said data processing step (2) by a learned processing method indicated by the processing method parameter.
A data processing method according to claim 9, comprising an input-data storage step (11) for storing a predetermined number of successive input data so as to provide said group of successive input data for said input-data evaluation step (12).
A data processing method according to claim 10, wherein, in said input-data evaluation step, the dispersion of the predetermined number of input data stored in said input-data storage step is calculated (32), and the degree of reliability (α_x) of the input data is calculated by forming the reciprocal of the dispersion.
A data processing method according to claim 10, wherein, in said input-data evaluation step (12), the average of the predetermined number of input data stored in said input-data storage step (11) is calculated, and the degree of reliability (α_x) of the input data is calculated according to the error of each datum against the average.
A data processing method according to claim 10, wherein, in said input-data evaluation step (12), the dispersion and the average of the predetermined number of input data stored in said input-data storage step (11) are calculated, the error of each input datum against the average is obtained, and the degree of reliability (α_x) of the input data is calculated according to the dispersion and the error.
A data processing method according to any one of claims 9 to 13, wherein, in said output-data evaluation step (13, 41), the current output data reliability is calculated according to the sum of the calculation of the current input data reliability and the calculation of the output data reliability obtained one datum before.
A data processing method according to any one of claims 9 to 14, comprising an output-data storage step (25) of storing the output data, wherein, in said data processing step (2), the previous output datum stored in said output-data storage step and the current input datum are added to obtain the output datum corresponding to the current input datum.
A data processing method according to claim 15, wherein the processing method parameter calculated in said real-time learning step (15) is a predetermined weight coefficient (w(t)), and output data corresponding to the current input data is obtained according to the weight coefficient in said data processing step (2).
A storage medium storing a computer-controllable program for processing successive input data and outputting the processed data, the program comprising the steps of a method according to any one of claims 9 to 16.
A computer program for processing successive input data and outputting the processed data, the program comprising the steps of a method according to any one of claims 9 to 16.